【作者】 施雪梅；

【导师】 赵耀；

【作者基本信息】 北京交通大学 ， 信号与信息处理， 2010， 硕士

【摘要】 计算机视觉作为一门新兴学科,它的研究目的是使用计算机代替人眼和大脑,根据观测到的图像对实际场景做出判断。图像序列中的运动目标检测与跟踪是计算机视觉领域研究的重要课题之一,它是计算机科学、人工智能、光学、数学等多学科的结晶,在导弹制导、工业产品探测、智能交通、人机交互领域具有非常重要的实用价值和广阔的发展前景。本文围绕图像序列的目标检测与跟踪技术,重点研究了可见光下的运动目标跟踪以及红外小目标检测等关键技术和方法。具体来说,论文的主要工作和贡献集中体现在以下几个方面：1.对于可见光下的运动目标跟踪,本文提出了一种基于贝叶斯分类的实时运动目标跟踪算法。算法以贝叶斯分类为核心,同时引入卡尔曼滤波算法预测目标位置,提高运算速度和定位精度。为了适应跟踪目标的尺度变化,文中采用双阈值机制对跟踪目标区域的统计特性进行判决,以此实现跟踪窗口的自适应调整。此外,本文还引入一种动态的模型参数更新策略来适应实际场景中目标和背景的变化。多个图像序列的实验结果表明了算法的鲁棒性和实时性。2.为了解决基于学习的目标跟踪算法中小样本及目标的高维表示问题,本文研究了将GLRAM(矩阵的广义低秩逼近)与PPCA(概率主成分分析)相结合用于运动目标跟踪的方法。该方法首先利用GLRAM算法对训练样本进行降维,获得目标图像的有效特征,然后根据降维后的训练样本建立PPCA模型。在后继帧中,直接利用概率模型进行判断,得到最优目标位置。考虑到实际目标在运动过程中的动态变化,算法中将新获得的目标加入到训练样本中,更新模型参数,提高了算法的鲁棒性。3.本文还对红外弱小目标检测的问题进行了研究,设计了一种基于时-空域联合的检测方法,来降低噪声对弱小目标检测的影响。通过对图像帧的差分和边缘检测运算,能够从不同角度得到目标的局部信息,将这些局部信息相结合,可以确定部分目标样本点,进而在边缘图像上寻找它们的连通域,即可获得完整的目标信息。同时为了提高目标检测的准确率,算法中还采用基于反馈机制的验证方法将目标检测和跟踪相结合,利用边缘和灰度信息来检验目标跟踪结果的准确性,有效降低了误检和漏检等情况的出现概率。更多还原

【Abstract】 Aiming at automatically understanding and analyzing the visual signals captured by various acquisition devices, computer vision has become one of the most advanced computer technologies and received a lot of attention in recent years. As an important subject of computer vision, object detection and tracking, an interdisciplinary field crossing computer science, artificial intelligence, optics and mathematics etc, has been widely utilized in a large number of practical cases, such as military visual missile guidance, industrial product detection and human-machine interface.The main goal of this thesis is to present some novel technologies to alleviate the key issues on object detection and tracking, and the elaboration of each technique is given as follows:1. To exactly track the interesting object in visible image, a bayesian classification based real-time object tracking method is proposed in this paper. Specially, the kalman filter has been explored to boost the tracking performance on computational complexity and localization precision. In terms of the constantly change related to the object scale, a dual-threshold mechanism based on the statistical analysis over the candidate object window was proposed to guarantee the adaptive adjustment of the tracking window. The model parameters have been also updated simultaneously according to the change of object and background in the practical scene. The experimental results on several video sequences show good real-time capacity and robustness of the proposed scheme.2. To alleviate the small samples size and high-dimension described problem in object tracking, an efficient method based on GLRAM(Generalized low rank matrix approximation) and PPCA(Probabilistic principal component analysis) has been presented in this thesis. First, the GLRAM has been utilized on the high-dimension features of training samples to obtain a compact representation, then, the PPCA model has been constructed based on them to seek the optimal object location of the candidate objects with the maximum PPCA model probability output in subsequent frames. In addition, taking account into robustness of the proposed technique to the change of the object, the PPCA model was updated dynamically in each frame.3. Aiming at alleviating the effect of noise on the performance of dim small object detection in infrared image sequences, a spatial-temporal based detection algorithm has been proposed by the utilization of the multi-view object information. The dim small object can be obtained by finding the connected component of these pixels with high confidence generated by frame difference and edge detection. Moreover, in order to improve the detection accuracy, the detection was combined with tracking process using validation approach based on feedback mechanism in which edge density and appearance information have been used to verify the accuracy of the results of tracking. This method reduced fall-out ratio and miss rate effectively.更多还原